Project Name

Vault template

First committed by asakin

Published in asakin/llm-context-base. Publisher page

74 GitHub starsUpdated 2026-10-25

What it is

A git template for building your own LLM-powered personal wiki. Training period, metadata standard, lint system included. Clone and go.

— the description by asakin

AutoGineer overview, from the brain’s own files

llm-context-base is an opinionated template designed to turn a local git repository into an LLM-powered personal wiki. It implements a three-layer pattern consisting of raw source documents, an LLM-maintained wiki, and a schema configuration. The system is designed to adapt to a user's workflow over time, eliminating the need for upfront taxonomy design. It includes a metadata standard for queryable files, an inbox-first capture system with a 7-day TTL, and a linting system to maintain document health. It supports various AI tools including Claude Code, Cursor, and GitHub Copilot, providing a consistent interface for managing personal knowledge.

How it works

Memory model: LLM wiki

AutoGineer overview, from the brain’s own files

The system uses a specific folder structure including 1-Projects, 2-Knowledge, 3-Journal, and 4-Private, alongside configuration and metadata folders. Users capture information into an _inbox folder, which the AI then processes and files according to a metadata standard. The system is bootstrapped via configuration files in .claude, .cursor, and .github, which direct the AI to follow instructions located in _config and _meta. It uses a training period to learn user patterns and maintains documentation through a split README file that separates user-specific content from framework documentation.

Folders
Folder	Role	Files
.claude	agent-config	2
.cursor	agent-config	1
.github	agent-config	1
.obsidian	obsidian-config	7
assets	attachments	2
docs	docs	13
examples	examples	4
1-Projects	folder	3
2-Knowledge	folder	5
3-Journal	folder	1
4-Private	folder	1
_config	folder	4
_inbox	folder	1
_meta	folder	14
_output	folder	1
_sources	folder	2

The agent reads

CONTRIBUTING.md evidence: .claude/CLAUDE.md
_config/config.md evidence: README.md

The agent writes

_config/config.md evidence: .claude/CLAUDE.md
_config/context.md evidence: .claude/CLAUDE.md
_inbox/ evidence: .claude/CLAUDE.md
_output/ evidence: README.md

Requirements

Apps: Obsidian

Databases: MongoDB, PostgreSQL

Works with: Claude Code, GitHub Copilot, Cursor, Devin (ex-Windsurf)

Set it up

Download the brain as a ZIP and unzip it where you keep your notes or project: https://mp-staging.autogineer.com/agent-brains/asakin/llm-context-base/download
In Obsidian, choose “Open folder as vault” and pick the folder.
Claude Code reads .claude/CLAUDE.md from this folder automatically.
cd llm-context-base && claude
AutoGineer scanned these files at commit 3c096ca.
First committed by asakin. Published in asakin/llm-context-base. Licence: Apache License 2.0.

Usage

AutoGineer overview, from the brain’s own files

Capture unstructured notes into the inbox for later organization by the AI.
Maintain decision records with context, options, and rationale.
Query the wiki to synthesize information across multiple documents.
Run linting commands to identify stale files or missing metadata.
Manage project initiatives with briefs and timeline tracking.
Generate reports or summaries from stored knowledge.

“Capture this: here's how our onboarding process works...”: Creates a file in the inbox and offers to file it in the appropriate location.
“We decided to use PostgreSQL over MongoDB, let me document why”: Generates a decision record with context, options, rationale, and an outcome section.
“What do we know about our hiring process?”: Scans file summaries and synthesizes an answer with source citations.
“Lint the wiki”: Surfaces stale files, missing metadata, and orphaned documents.

Documentation by asakin

[Project Name]

This README will be personalized during your first session.

llm-context-base

An opinionated template for building your own LLM Wiki.

Use the template, point your AI assistant at it, and start talking. Over time, it learns how you work and adapts its structure to your needs without you having to design anything upfront.

How This Relates to Karpathy's LLM Wiki

Andrej Karpathy's LLM Wiki gist describes a three-layer pattern: raw source documents → an LLM-maintained wiki → a schema config file (like CLAUDE.md or AGENTS.md). The LLM reads sources, maintains structured wiki pages, and follows the schema config to stay consistent.

llm-context-base implements this pattern as a ready-to-use template — so you don't have to build it from scratch. What it adds on top of the base pattern:

Metadata standard — every file gets a queryable header (type, summary, tags, status, date). The LLM scans these instead of loading full files, which scales cleanly to hundreds of documents.
Inbox-first capture — a frictionless landing zone with a 7-day TTL. You capture without deciding where things go. The LLM files them later.
Training period — the system starts chatty (learning your patterns) and goes quiet as it adapts. You never design a taxonomy upfront.
Lint system — surfaces stale files, missing metadata, and orphaned docs on demand.
Multi-LLM support — bootstrap shims for Claude Code, Cursor, Copilot, Windsurf, and others. One config, any tool.

Full comparison with the Karpathy pattern →

Quick Start

1. Create your copy. Click "Use this template" on GitHub. Make it private. Clone it.

git clone https://github.com/YOUR-USERNAME/my-wiki.git
cd my-wiki

2. Set up your workspace. Open the repo as an Obsidian vault on one side, your AI agent on the other. You talk to the AI — it builds the wiki. You watch it happen in Obsidian.

Full Obsidian setup → · No Obsidian? Any editor + any supported AI tool works fine.

3. Fill in your profile. Open _config/config.md, fill in the About You section, and set Training Start Date to today.

4. Start talking. Open in your AI tool of choice and go:

"Capture this: here's how our onboarding process works..."
"We decided to use PostgreSQL over MongoDB, let me document why"
"What do we know about our deployment process?"
"Lint the wiki, what needs attention?"

That's it. The system adapts to what you actually need.

Want to stay updated as the framework improves? The template approach gives you a clean start but disconnects you from upstream. If you want to pull future improvements (new protocols, better training behavior, bug fixes), use the Framework Track setup instead.

What Makes This Different

It learns how you work. The first 30 days are a training period. Your AI asks questions, suggests directories, and logs what it learns. By day 45, it's quiet and efficient. You never design a taxonomy upfront. Full details →

Capture first, organize later. Everything lands in _inbox/ with a 7-day TTL. You never hesitate to capture because you don't have to decide where it goes yet. Your AI files it later.

Your README stays yours. This file has two sections split by a marker. Above it: your personal project README, maintained by your AI as the wiki evolves. Below it: the framework docs, which track upstream cleanly. You get living documentation without merge conflicts.

Patterns are captured as they emerge. When your AI learns a new convention or preference during a conversation, it writes it down immediately, not at commit time. Nothing is lost if you commit from the command line without AI involvement.

Every file is queryable. A metadata standard on every document means your AI can scan summaries and tags without loading full files. This scales to hundreds of documents.

It maintains itself. Lint checks surface stale files, missing metadata, and orphaned docs. The system even reviews its own instruction efficiency to prevent context window bloat as your wiki grows.

What You Can Do

Say this	What happens
"Capture this: here's how we set up monitoring..."	Creates a file in `_inbox/`, offers to file it
"We decided to use Terraform over Pulumi, document why"	Decision record with context, options, rationale, and an outcome section to fill in later
"What do we know about our hiring process?"	Scans summaries, synthesizes an answer with source citations
"Lint the wiki"	Surfaces stale files, missing metadata, orphaned docs
"Start a new initiative: migrate to Kubernetes"	Project workspace with brief, timeline, artifact tracking
"Create a slide deck summarizing Q1 decisions"	Generates output in `_output/`, dated and organized

These form a cycle. Capture feeds queries, queries inform decisions, decisions get tracked, lint keeps it all healthy. The system compounds over time.

Go Deeper

Topic	What's there
Training Phases	How the system adapts to you over 30 days
Directory Structure	What each folder is for, why there are only 4
Journal Sync & Privacy	Syncing `3-Journal/` across devices — plaintext, git-crypt, or external
Metadata Standard	The format that makes everything queryable
Supported Tools	Claude Code, Cursor, Copilot, Windsurf, ChatGPT, and others
Obsidian Setup	Pre-configured vault, plugins, web clipper
Design Decisions	Why the system works the way it does
FAQ	Common questions
Linked Code Repos	Connect wiki project entries to real code repos — workspace files, context files, Obsidian plugin dev
Extending & Contributing	Customize locally vs. contribute upstream
Future Directions	Where this is headed
Philosophy	The deeper thinking behind the system

The examples/ directory has four sample files showing the metadata standard in action. Read them, then delete them when you're comfortable.

Credits

Built on Andrej Karpathy's LLM Wiki Pattern. Implementation by Ariel Sakin, refined over several months of daily use managing business operations.

License

Apache 2.0. See LICENSE.

docs/obsidian-setup.md

Using with Obsidian

The recommended way to use this system: Obsidian on one side, your AI agent on the other. You talk to the AI. It reads, writes, and organizes your wiki. You watch the results appear in Obsidian in real time — follow links, check the graph view, browse what's changed.

Obsidian is where you browse your wiki. You can talk to your AI directly inside Obsidian using a plugin, from a terminal alongside it, or from any AI tool you already use — Claude Code, Cursor, the Claude app, whatever you prefer.

Getting Started

1. Open the vault. This repo ships as a ready-to-use Obsidian vault. Open the cloned directory as a vault in Obsidian — graph colors, core plugins, and display settings are pre-configured.

2. Open your AI tool next to it. No plugin needed. Just open Claude Code, the Claude desktop app, Cursor, or ChatGPT alongside Obsidian. You talk to the AI in one window, the wiki updates in real time in the other.

That's the whole setup. Obsidian is the viewer. The AI is the writer. They don't need to be connected — they both just read and write the same folder of files.

3. Start talking. See the Quick Start for what to say first.

What's Pre-configured

Already set in the shipped vault, no action needed:

Setting	Value	Why
New files default location	`_inbox/`	Ctrl+N always creates in the inbox. Inbox-first without thinking.
Templates folder	`_meta/templates/`	Ctrl+T inserts from the included set.
Daily notes location	`3-Journal/`	Daily note button creates a journal entry with the right template.
Attachments location	`assets/`	Images stored in a fixed directory, not scattered.
Graph view	Color-coded by directory	Projects (blue), Knowledge (green), Journal (purple), Private (red), Inbox (orange), System files (gray)
Workspace files	Gitignored	Your local layout stays private.

Optional Plugins

All plugins here are optional. The system works without any of them.

To install community plugins: Obsidian Settings → Community plugins → Turn off restricted mode → Browse.

Want AI inside Obsidian instead of a separate window?

By default you use an AI tool (Claude Code, Claude app, Cursor, ChatGPT) in a separate window next to Obsidian. If you'd rather type to the AI inside Obsidian itself, these plugins add that:

Plugin	What it does
Claudian	Claude Code panel inside Obsidian — runs your full CLAUDE.md session protocol
claude-obsidian	Claude Code with hot cache for faster context loading — same as above, faster startup
Terminal	Embedded terminal pane inside Obsidian — run `claude` directly without switching apps
Copilot	AI chat sidebar — works with Claude, GPT-4, and others. Note: this is a generic chat panel and does not run your CLAUDE.md session protocol — your instruction modules, training phase, and inbox behavior won't apply

Want to save web pages directly to your inbox?

Install the Obsidian Web Clipper browser extension. See the Web Clipper Setup section below for the template.

Other useful plugins

Plugin	What it does
Dataview	Query engine for metadata — build dashboards and filtered views
Kanban	Visual kanban boards backed by markdown
Periodic Notes	Daily/weekly/monthly notes mapped to `3-Journal/` — extends the built-in daily notes with weekly and monthly views

Dataview caveat: This system's metadata format (**Type:** knowledge, single colon) is incompatible with Dataview's inline field syntax (**Type**:: knowledge, double colon). Dataview cannot query your existing notes as written. To use Dataview, you either need to migrate your files to double-colon syntax or switch to YAML frontmatter — both are significant changes to how the system operates. If dashboard queries are important to you, use YAML frontmatter from the start. If not, skip Dataview.

Web Clipper Setup

If you install Obsidian Web Clipper, import the pre-built template to drop clips straight into your inbox with the right metadata:

Open the Web Clipper extension → click the gear icon → Templates
Click Import and select assets/clipper-template.json from this repo
Set it as your default template

That's it. Every clip lands in _inbox/ with the metadata standard applied.

The template adds one important field:

> Why I saved this: [fill in one sentence before closing the clipper]

Fill this in before you close the clipper — one sentence is enough. Your AI uses it at triage to decide what to do:

What you write	What the AI does
"read later"	Leaves it in inbox, surfaces it at next session
"save the key points on X"	Extracts the relevant knowledge, files it, discards the raw clip
"relates to [project name]"	Links it to or files it under that project
"source for [draft name]"	Pulls relevant quotes into the draft
(left blank)	Asks you what you want before doing anything

If you skip the line, your AI will ask before filing — it won't guess.

Tips

Store images locally. Bind "Download attachments for current file" to a hotkey (e.g., Ctrl+Shift+D). The vault saves attachments to assets/. This downloads remote images to that local folder, preventing broken URLs. (Tip from Karpathy's original post)

Use the graph view. As your wiki grows, the graph view becomes a map of your thinking. Color-coded directories make it easy to spot clusters and gaps.

Let the AI handle organization. Resist the urge to manually file things. Capture into _inbox/, then let your AI suggest where it goes. That's the whole point.

Making It Look Good

The vault ships with a CSS snippet (llm-wiki.css) and graph view settings pre-configured. But you need to install a theme to get the full effect.

Step 1: Install the Minimal theme (recommended)

Open Obsidian → Settings → Appearance → Themes → Browse
Search for Minimal and install it
The vault's appearance.json already points to it — it activates automatically

Minimal is the most popular Obsidian theme. It has clean typography, good dark/light mode support, and a companion plugin (Minimal Theme Settings) for further customization without touching CSS.

Adapt to system theme: In Settings → Appearance → Base color scheme, set it to Adapt to system. Obsidian will follow your OS's light/dark schedule automatically.

Step 2: CSS snippet (already active)

The vault ships with llm-wiki.css in .obsidian/snippets/. It's already enabled. It does:

Metadata card — the **Type:**, **Summary:**, etc. block at the top of every file renders as a subtle card with an accent border, not raw bold text
Typography — readable line length (760px max), better heading hierarchy, 1.75 line height
Tables — cleaner borders, hover highlight
Blockquotes — styled as pull quotes with an accent bar

If you want to customize it: Settings → Appearance → CSS snippets → pencil icon next to llm-wiki.

Step 3: Graph view

The graph is pre-configured with colors per directory:

Color	Directory
Blue	1-Projects
Green	2-Knowledge
Purple	3-Journal
Red	4-Private
Orange	_inbox
Gray	_meta, _config
Teal	_output, _sources

If Obsidian resets your graph settings (it sometimes does when the vault reopens), you can restore them by running git checkout .obsidian/graph.json and restarting Obsidian.

73 files

Advantages

The system adapts to user patterns over a 30-day training period without requiring upfront taxonomy design. evidence: README.md
Metadata standards on every document allow the AI to scan summaries and tags without loading full files, enabling scalability. evidence: README.md
The inbox-first capture system allows for frictionless entry of information with a 7-day TTL for later filing. evidence: README.md

Limitations

The template approach disconnects the user from upstream updates unless the Framework Track setup is used. evidence: README.md

Security

Risk level: No risk found

This brain runs

Agent instructions that run or fetch something (1)

.claude/CLAUDE.md: **Exception:** If the user's first message is a `/init` command or a meta question about the repo setup, complete that first — then run the Session Start Protocol on the next exchange.